Printable math worksheets, sized to a five-minute drill

Free printable math practice for roughly 3rd–5th grade — word problems, logic puzzles, and arithmetic drills, each one page per kid plus a worked answer key.

→ math.sifxtreme.com — print them, no signup, no ads

The site also has step-through animations of the written methods — long division and multiplication, one pencil mark at a time, with the trap called out at the step where it actually bites. This repo is the generator behind it all: self-contained HTML, no dependencies, python3 and a browser.

Two things here are unusual and worth the thirty seconds:

• Drills are sized to a clock, not a page. Most worksheet sites give you 20 or 25 problems whatever the skill. 20 long divisions is a sixteen-minute sheet. If you want five minutes, the count has to come from the work, and that's what drill_cost.py computes.
• The grid forces the work to be shown. Every step of the written method gets its own printed box, so a child can't jump to an answer. That rule exists because one of them did exactly that. It's optional — --layout plain gives the same problems with blank space, for when the method is solid and you want it the way a test would ask.

---

Setting it up for your kids

One step. Copy the example config and put your own names in it:

cp kids.example.json kids.local.json
$EDITOR kids.local.json

{
  "kids": [
    { "id": "kid1", "name": "Amira", "grade": 3 },
    { "id": "kid2", "name": "Idris", "grade": 4 }
  ]
}

That's it — generate a sheet and the names are on it.

kids.local.json is gitignored. kids.example.json is not. Committed code refers to children by id (kid1, kid2) and never by name; the name is read at render time and appears only in the printed Name: field. So you can share, fork, or publish the repo without shipping your family with it.

If you skip this step nothing breaks — the generator falls back to the example file and prints a loud notice. Sheets will say "Kid One". That's deliberate: a silent fallback would hand you a stack of paper addressed to nobody.

More or fewer than two kids

ids are the contract. The drill plans in drill_gen.py are keyed on kid1 and kid2:

PLANS = {
    "div": [("kid1", (2, 1), "2-digit ÷ 1-digit, with a remainder", "div_2d_1"),
            ("kid2", (3, 1), "3-digit ÷ 1-digit, with a remainder", "div_3d_1")],
    ...
}

Add a kid3 to your config and a matching row there. The digit shapes (2, 1) mean 2-digit dividend ÷ 1-digit divisor; the last field picks the time model from drill_cost.py.

What is and isn't in this repo

Being straight about it, because a half-private repo is worse than a known-public one. This repo is public, and it was scrubbed deliberately before it was published:

✅ All code	keyed on kid1/kid2 ids, no names
✅ private/ and kids.local.json	gitignored — the practice plan, results log, and any per-child notes live there
✅ Generated sheets	gitignored; rebuilt from a seed
✅ Every Name: header field	blank — the child writes it
✅ Git history and commit messages	rewritten with git-filter-repo; no names, no school, no per-child records
ℹ️ Story characters	The word problems and logic puzzles are cast with Muslim names — Idris, Hamza, Bilal, Amina, Zayd, Tariq and others. Those are characters, kept on purpose: the puzzle answers reference them by name. Nothing indicates whose repo it is.

If you fork this for your own kids, kids.local.json is the only file that ever holds a real name, and it is gitignored from the start.

---

Daily use

A full day is three sheets per kid: word problems, logic, and a drill.

# drills — generated, sized to five minutes
python3 drill_gen.py --skill div --seed $(date +%Y%m%d)   # or --skill mul
python3 drill_gen.py --skill zeros --stage 1             # kid1 Noetic prep: 2-minute trailing-zero probe (stages 1|2|3|mix)
python3 verify_drills.py                                  # must exit 0
./print-worksheet.sh worksheet-division-drill.html --dry-run

# word problems / logic — pick the day's sheet from PRACTICE-PLAN-2026.md
python3 build_focus.py && python3 verify_focus.py       # Noetic-prep focus-track word sheets (one kid per spec; outputs gitignored)
python3 build_fracmul.py && python3 verify_fracmul.py   # kid2 fraction packet: 10 multiply + 10 add/subtract, student + key files
./print-worksheet.sh worksheet-sailing-ships.html --date today

--budget 600 makes it a ten-minute sheet instead; the problem counts re-derive themselves. --seed makes a sheet reproducible — same seed, same problems.

⚠️ print-worksheet.sh and print-drill.sh refuse to print ahead, print a batch, or reprint a day that already exists. That's day-guard.sh, and it is deliberate. A genuine replacement goes through --override-day-guard, which is loud and logged.

The files

File	What
drill_gen.py	generates division/multiplication drills, sized to a time budget
drill_cost.py	the time model — marks per problem → seconds → problems per sheet
verify_drills.py	audits the rendered HTML independently. Run before printing
build_sheets.py	generates the word-problem and logic sheets from specs_*.py
verify_sheets.py	re-derives every word-problem answer. Must exit 0
build_focus.py / specs_focus.py / verify_focus.py	one-kid Noetic-prep focus sheets; student sheet and key are separate files. The verifier finds each answer by brute force and must exit 0
build_fracmul.py / specs_fracmul.py / verify_fracmul.py	kid2's 20-problem fraction packet (10 multiply, 10 add/subtract), student and key files. The verifier recomputes every answer with Fraction and must exit 0
print-worksheet.sh	render to a clean PDF and print
print-drill.sh	same, for an external drill PDF
day-guard.sh	the one-day-at-a-time print guard
test-day-guard.sh	unit tests for the guard's logic — calls guard_one_day directly
test-print-path.sh	end-to-end test of the whole print path, via a stub lp. Catches what the unit tests structurally cannot
animations/	step-through animations of the written methods
AGENTS.md	the real documentation — rules, difficulty pin, hard-won lessons
DRILLS.md	external drill sources (fallback)
KUMON.md	the Kumon sample packets — what's in them, and how long a page is supposed to take
kumon_cost.py	prices each Kumon page with the drill_cost.py model, against Kumon's own 15–30 min

AGENTS.md is the file to read before changing anything. It carries the rules that have already decayed once and been reinstated — why drills are one day at a time, why the work has to be shown, and why a doc that says "never" needs its date checked.

The public site — math.sifxtreme.com

A scrubbed subset of this repo is published as a free printable-worksheet site. It is a separate Cloudflare Pages project (sifxtreme-math) from the blog, on purpose: a bad math deploy can then never touch sifxtreme.com.

python3 build_site.py          # renders site/ from the SHEETS allow-list
python3 verify_public.py       # MUST exit 0 — refuses to publish if anything private survived
export CLOUDFLARE_API_TOKEN=$(grep ^CLOUDFLARE_API_TOKEN= ~/code/experiments/cloudflare-cli/.env | cut -d= -f2-)
npx wrangler pages deploy site --project-name sifxtreme-math --branch master

⚠️ If the token looks dead, check with cf, not curl. This site shares its credential with ../typing/. A cfat_ token is an account token and returns Invalid API Token from /user/tokens/verify — the user endpoint — however healthy it is; wrangler misleads identically by failing on /accounts. On 2026-08-10 that cost an hour and an invented expiry story for a token that was alive and merely lacked Pages permission. ../../cloudflare-cli/cf whoami reports it correctly. Full write-up: ../typing/AGENTS.md § Deploying.

⚠️ Verify by grepping the live page, never by status code — Pages serves its fallback with a 200, so an unpublished URL still looks fine:

curl -sL https://math.sifxtreme.com/ | grep -o '<title>[^<]*</title>'

What makes a sheet publishable: the SHEETS allow-list in build_site.py. A new worksheet is private until someone adds it — the safe default. verify_public.py is the last line, not the only one, and it reads the terms it hunts from the gitignored private/forbidden-terms.txt rather than hardcoding them.

Calibrating the time model

Every problem count is currently a prediction. The model is anchored on one real measurement (49 problems of 3-digit subtraction in 13 minutes) and extrapolated from subtraction to division and multiplication.

To make it a measurement instead: time a real sheet, then

import drill_cost
drill_cost.calibrate(observed_sec=41.0, marks=15.0)   # → the SEC_PER_MARK it implies

and update SEC_PER_MARK. Don't hand-adjust the problem counts — that breaks the link between the model and reality, and then the numbers mean nothing.